apply crop and flip transformations to bboxes and labels

[bernardev254]

Summary

This PR completes the implementation of bounding box transformation logic in the create_global_crops() function in hackathon/augmentations.py.

Changes

• Applies random_resized_crop and random_horizontal_flip to images and bounding boxes.
• Filters out invalid (fully cropped out) bounding boxes.
• Extracts corresponding labels using valid indices.
• Stores per-crop detection targets as dictionaries with "bboxes" and "labels" under crops_heatmaps.
• Keeps the original output structure to avoid major changes in the data pipeline.

Motivation

Bounding box augmentation was previously unimplemented (NotImplementedError). This PR ensures that
global crops of images are now paired with appropriately transformed bounding boxes and class labels, enabling object detection models to learn from augmented data.
closes #2